fix(serve): delete two lib tests that assert nothing and take 19 minutes; scale four more

[noahgift]

These were the top cause of workspace-test timing out at its 75-minute step limit — which blocks every PR in the repo. They were invisible until #2517 fixed nextest's suppressed SLOW reporting; before that the job died anonymously.

From CI job 95324516796. "Max threshold" is the highest SLOW [> Ns] line seen, so real durations are higher — and several appeared as TRY 2 SLOW, i.e. being retried (retries = 2), tripling the cost:

max threshold	test	asserts
>1140s (19 min)	test_parity017e_end_to_end_batch_throughput	0
>300s	test_parity017a_gpu_batch_ffn_implementation	1
>300s	test_parity018b_batch_ffn_gpu_method	1
>120s	test_parity015d_batch_forward_timing	0

017e allocated batch 32 × hidden 2560 × intermediate 10240 × 4 layers of f32 — gigabytes — measured wall-clock with Instant, printed a throughput number, and asserted nothing. On a CPU-only runner the "GPU" path falls back to CPU. A 19-minute test that cannot fail is the purest form of the assertions-must-exclude-an-outcome defect.

Deleted (0 asserts), with tombstones pointing throughput work at benches/

Coverage lost, stated precisely rather than as "none":

• 015d: nothing. It did let _ = scheduler.matmul(...) — result discarded. It could not fail even if matmul returned Err.
• 017e: not literally nothing. Zero asserts but two .expect() calls, so it did check Err-freedom for a 4-layer FFN at phi-2 sizes. That same Err-freedom is still covered by 017a/018b on the same code path; what's genuinely gone is Err-freedom at phi-2 tensor sizes specifically, plus the 4-sequential-layer loop. A fair trade, but not zero.

Scaled (assertion kept)

017a, 018b, plus 015a and 016c found by enumerating the family programmatically. batch 32→8, hidden 2560→128, intermediate 10240→512 — 1600× less arithmetic. Dimensions chosen so m*k*n = 524,288 stays above HybridScheduler's gpu_threshold (262,144), so the GPU-vs-CPU dispatch decision is unchanged (verified: Should use GPU: true still prints). Meaningless Instant/GFLOPS prints removed; no timing assertions present or added.

A real defect found while mutation-testing — worth more than the speedup

Mutating HybridScheduler::matmul to drop one output element left 017a green. Its Err arm printed "SKIP - GPU path failed" and passed. Same hole in 015a and 016c:

-  println!("  Error: {} (expected if no GPU)", e);
-  println!("  Status: SKIP - Matmul failed (may be CPU fallback)");
-  println!("  Status: SKIP - GPU path failed");

Three tests passing when the operation they test failed — and the reassurance was false: cpu_matmul returns Ok unconditionally, so "may be CPU fallback" could never explain an Err.

Proven pre-existing rather than caused by the scaling, by re-running the identical mutation at the original [32×2560] dimensions — also test result: ok. Those arms are now panic!, after which all 8 combinations (4 tests × GPU and CPU-forced paths) go RED, rc=101.

Measurements

CPU-only path forced with VK_ICD_FILENAMES=/nonexistent and proven engaged (GPU available: false, GFLOPS 9.99 → 0.25) rather than assumed:

test	before	after
017e	55.3s	deleted
015d	6.5s	deleted
017a	13.9s	0.03s
018b	13.9s	0.03s
015a	7.0s	0.02s
016c	7.0s	0.02s
total	103.6s	0.10s

Absolute seconds here run ~20× below CI's for both 017a and 017e — a consistent ratio, being runner concurrency contention across 15,721 tests. Treat the ~1000× reduction as the reliable figure, not the absolute times.

Verification

cargo test -p aprender-serve --lib	rc=0, 15,660 passed, 0 failed, 59 ignored
test count	15721 → 15719, exactly the two deletions
cargo fmt --check	rc=0
cargo clippy --all-targets	rc=101 — pre-existing, proven by materialising origin/main's version of these 5 files via git show and getting the identical rc and same 4 errors in files never touched here. No lint config changed.

_{CI cannot currently verify this: the fleet is down again (#2528, reopened — rustup deleted a second time). Everything above is local.}

Refs #2503, #2517